MNN is a lightweight C++ inference engine for on-device ML, offering multiple backends and tooling for model conversion, quantization, and deployment. The project has extensive release history and documentation, with active maintenance as of 2026-08-04.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
MNN is a highly efficient and lightweight deep learning framework. It supports inference and training of deep learning models and has industry-leading performance for inference and training on-device.
How it works
Key features include support for multiple backends (CPU, GPU, NPU) and various architectures; quantization (FP16 / Int8) and multiple device targets (iOS, Android, embedded). The project provides tools such as MNN-Converter, MNN-Compress, MNN-Express, MNN-CV, and MNN-Train to handle model conversion, optimization, and deployment.
Getting started
Install guidance is available via releases:
- 3.6.1:
pip install mnn==3.6.1 - 3.6.0:
pip install mnn==3.6.0 - 3.5.0: (not fully shown in excerpt beyond description)
Recent releases
Latest releases shown in the excerpt:
- 3.6.1 (2026-07-23): Hexagon backend, Transformer C4 Fuse backend performance optimization. Installation:
pip install mnn==3.6.1. - 3.6.0 (2026-06-16): Installation:
pip install mnn==3.6.0. - 3.5.0 (2026-04-07): Multiple backend LLM inference capabilities, high-performance quantization, edge speech enhancements.
- 3.4.1 (2026-03-05): Qwen3.5 support and Linear Attention, LLM resource management and security fixes.
- 3.4.0 (2026-02-07): GPU/QNN backend deepening, Attention memory optimizations, and GPU stability.
Traction
- Stars: 15808 (from repo data).
- Forks: 2394
- Open issues: 78
Behind the repo
No specific linked startup or company page is provided in the given facts beyond Alibaba origin and product integration into Alibaba apps.
Caveats
License: Apache-2.0. Created: 2019-04-15. Last push: 2026-08-04.